Maturity Active development 33 of 120 roadmap tasks complete (1 of 6 phases). Not a released product.
Quality and review
Six evaluators inspect a draft and return typed findings. None of them approves anything, and the difference is the whole design.
Five run as structured model calls inside the generation graph. Chrono makes no model call at all.
Evaluators execute in parallel over a draft and aggregate before the graph routes on the result. Each returns typed findings and, where a model was involved, normalized usage metrics for cost accounting. Schemas are versioned alongside the prompts and the orchestration code.
Factual accuracy and claim support.
Spoken runtime estimation.
Overused tropes, cliches, and snowclones.
Pronunciation gaps.
Brand-guideline compliance.
False endings.
Two of six are implemented. The rest are specified with their contracts and not yet built; the status on each record says which is which.
Structured, per-claim, and carrying remediation guidance rather than a verdict. The example is synthetic.
{
"claim_id": "c-014",
"claim_kind": "attributed-quotation",
"cited_sources": ["src-transcript-01"],
"support_level": "partial",
"severity": "high",
"remediation": "The transcript records this phrasing at 00:14:22 but attributes it to the second speaker. Confirm the attribution or cut the quotation."
}
A finding names what it examined, what it could and could not tie to a source, and what a person might do about it. It does not say the script is wrong, and nothing downstream treats a severity as an instruction.
Stated as a matrix. Status is per row, because the boundary is delivered in pieces.
| Activity | The system | A person | Status |
|---|---|---|---|
| Weighting and selecting between conflicting sources | Computes priority scores from quality, freshness, and reliability, and applies the weighting matrix. | Sets the weighting coefficients per series profile, and can inspect every rejected source. | Available |
| Estimating episode runtime | Produces a deterministic estimate and records the estimator metadata behind it. | Decides whether the estimate is close enough to commission the recording. | Available |
| Scoring a draft for accuracy | Returns per-claim findings with severity and remediation guidance. | Decides whether a finding is correct and what to do about it. A score is not a verdict. | In progress |
| Spending money on model calls | Reserves budget before a billable call, commits on success, and releases on failure. | Sets the per-episode budget and the tier of model each task may use. | In progress |
| Approving an episode for an audience | Surfaces findings, suspends the graph, and preserves state while the decision is pending. | Approves, requests changes, or redirects the run. This decision has no automated path. | Planned |
Two mechanisms stop a run: an editorial guardrail and an exhausted budget.
Guardrails persist on series profiles and episode templates, so a generation request carries stable editorial instructions as its system prompt rather than depending on whoever composed the call. Changing the guardrail changes every subsequent run from one place, under the same change history and optimistic locking as the rest of the profile.
Budgets use reservation semantics around each billable call: reserve, then commit on success or release on failure. A run that cannot reserve does not make the call. Cost line items pin the provider pricing snapshot in force at the time, so a later price change does not rewrite what an episode cost.
Designed, not delivered. The honest state of the most important part of the system.
A checkpoint suspends the generation graph and preserves its state while an editor considers the findings. Because the state is durable, a review period can last as long as it needs to without losing the run - the graph resumes rather than restarting.
The reviewer actions the design provides for are approve, request changes, and redirect the run with targeted remediation prompts. Drafts that fall below threshold re-enter generation; drafts above it proceed to approval. Which threshold, and whether a team wants automatic re-entry at all, is a series-level decision.